INDEX / COMPARE / DOC UH-26-018

Four Archetypes Behind the "Undressher AI" Keyword

Services answering to this search term cluster into four recurring shapes. They differ in delivery mechanism — not in what they do to a real person's photo. Scores below run 0–10 where 10 is the highest observed risk.

By Dr. Maren I. Voss, independent ML-audit researcher · Last updated: November 2026 · Reading time ~6 min

C-1Scoring table

ArchetypeTypical formRetention riskOperator accountabilityChurn rateAggregate risk
Web upload toolSingle-page site, card or crypto checkout9/109/10 noneHigh9.0
Telegram botBot handle, crypto or "credits"9/1010/10 noneVery high9.3
Mobile appStore listing, in-app purchase7/105/10 someModerate7.0
Local open-source modelUser-run checkpoint, no upload2/10n/aLow5.5*

*Local execution removes the retention path but not the legal exposure — creating the image is still the offence in many jurisdictions. Score reflects legal + ethical residue, not data theft.

C-2Archetype notes

Web upload tools

The dominant shape for this keyword: a landing page, an upload box, a queue timer, a paywall. Infrastructure sits behind Cloudflare, the operator behind a privacy-guarded registration. The business model is retention — free tiers exist to collect inputs, paid tiers to monetize impatience. Expect the domain to be under a year old and the privacy policy to be boilerplate. Evaluation steps 2–4 of the main protocol surface this within minutes.

Telegram bots

Where services migrate after domain seizure. Discovery happens through invite links and channel networks rather than search. Payment is crypto or "credit" packs sold by username. From a research standpoint, bots are the least auditable archetype — no public policy page, no domain records, no payment-processor leverage point for regulators.

Mobile apps

Periodically flood app stores under "face editor" or "art filter" branding before removal waves. Slightly more accountable — store billing creates a paper trail — but the upload still leaves the device and store delisting does not recall installed copies. Documented removals ran in waves through 2024–2025 after press investigations.

Local open-source models

Fine-tuned checkpoints circulating on model hubs and forums. Technically distinct: no upload leaves the machine, so the retention risk disappears. What does not disappear is the creation offence — several jurisdictions now criminalize generating the image at all. Hosting platforms have removed such checkpoints in sweeps, pushing distribution to torrents and private Discords.

Field noteThe comparison that matters is not tool-vs-tool but claim-vs-evidence. Every archetype claims "privacy" and "auto-delete"; none has ever published an audit supporting it. Score the evidence, not the adjective.

C-2bMethodology behind the scores

Each score derives from the eight-step protocol on the main page, run against a representative sample of live services per archetype during October–November 2026. Retention risk rates the likelihood an uploaded photo persists beyond the stated session: unverifiable "auto-delete" claims score as retained. Operator accountability scores the presence of a named legal entity, a working corporate registration, and a reachable abuse contact — three checks no web tool or bot in the sample fully passed. Churn rate reflects observed domain lifespans in the logged set: median five months for web tools, shorter for affiliate mirrors and bot handles. The aggregate is a straight average, not a weighted one — these archetypes fail so broadly that weighting would only disguise it.

Two deliberate exclusions: output realism (irrelevant to risk scoring — a more convincing fake is more harmful, not less) and price (all archetypes cluster at the same few-dollars-per-generation range).

C-3Legitimate contrast

Worth stating: none of the four archetypes is a privacy-respecting alternative to the others, because the underlying act — synthesizing intimate imagery of real people without consent — is the problem. The legitimate tools in adjacent space work on consent-bound material (self-directed avatars, licensed fashion try-on) and look different in every measurable way: named companies, real privacy terms, provenance-signed outputs, and no anonymous payment rails.

Documentation index